2026-06-18 — views

AV Night and Low-Light Performance — How Camera-Only and LiDAR Stacks Handle Darkness

Night and adverse weather cover ~50% of real driving risk — how Tesla camera-only and Waymo LiDAR stacks perform determines AV geographic and commercial scale.

Article 53 in the Physical AI Benchmark Series — Darkness, Rain, and Glare

Operational design domains tell you where an autonomous vehicle is allowed to go. Night and adverse weather tell you where it will actually scale. These two conditions — darkness and precipitation — are not edge cases. Approximately 25% of all US vehicle miles are driven between 9 PM and 6 AM (NHTSA est.), yet roughly 50% of fatal crashes occur during those same hours. Rain, fog, and snow are implicated in approximately 17% of all vehicle crashes (FHWA). An autonomous vehicle that cannot reliably operate in low light and wet conditions can serve only a fraction of the total addressable driving demand.

The most persistent technical debate in autonomous vehicles — Tesla’s camera-only approach versus Waymo’s LiDAR-plus-camera-plus-radar stack — is ultimately settled not in daylight suburban scenarios where both systems perform well, but in the hard conditions that expose architectural limits. This article maps those limits: what each sensor sees in darkness, how each system degrades under rain and fog, what both companies have done operationally to manage those limits, and what must change before either stack can serve snow-belt geographies at commercial scale.

All figures marked (est.) are estimates based on published research, public company disclosures, and industry reporting. Sensor performance figures have not been independently verified under controlled test conditions.

Section 1 — Why Low-Light Performance Is an AV Ramp Gate

Human drivers are approximately 3x more likely to be involved in a fatal crash at night versus daytime (NHTSA est.). For an autonomous vehicle to be net-safer than human drivers — the threshold that justifies commercial deployment and regulatory approval — it must outperform that 3x-worse-at-night human benchmark, not merely match it.

Metric	Detail
Share of US miles driven at night	~25% of all US miles between 9 PM and 6 AM (NHTSA est.)
Share of fatal crashes at night	~50% of fatal crashes occur at night despite 25% of miles (NHTSA)
Human night driving risk multiplier	~3x more likely to be involved in a fatal crash at night vs. daytime (NHTSA est.)
AV safety threshold	Must outperform 3x-worse-than-daytime human benchmark, not merely match it
Adverse weather crash share	~17% of all vehicle crashes involve adverse weather (FHWA)

The implication for commercial scale is direct: a robotaxi that pulls off the road after dark or refuses service in rain is not competing with human drivers — it is a fair-weather supplement. Geographic expansion into dense northeastern US cities, European markets, or Asian markets requires operating through rain, glare, and at minimum urban night conditions. Snow-belt performance is the furthest horizon and currently beyond both major stacks.

Section 2 — Sensor Physics: What Each Technology Sees in the Dark

Understanding the debate requires understanding the physics. Different sensors have fundamentally different relationships with ambient light.

Sensor	Night performance	Adverse weather performance	Key limitation
Visible-light camera	Degrades significantly in darkness; depends on streetlights and headlights	Rain on lens degrades image; fog scatters light severely	Passive sensor — only sees reflected or emitted light
Near-infrared (NIR) camera	Better than visible in low ambient light; used in some pillar cameras (est.)	Similar rain and fog limitations	Still passive; limited effective range at night
LiDAR (spinning or solid-state)	Does not depend on ambient light; active laser source; performs equally day and night	Rain and snow scatter laser returns; heavy precipitation degrades range	Expensive; loses range in heavy precipitation
Radar (millimeter-wave)	Nearly unaffected by darkness	Penetrates rain, fog, and snow; best adverse-weather sensor available	Low spatial resolution; cannot identify object class, color, or texture
Thermal infrared camera	Detects heat signatures (pedestrians, animals) in total darkness	Less affected by rain and fog than visible cameras	Expensive; no color; limited range

The fundamental divide is active versus passive. Cameras are passive sensors — they record photons that exist in the environment. At night, in the absence of artificial light, a camera has almost nothing to record. LiDAR and radar are active sensors — they emit their own energy (laser pulses, radio waves) and record the return. Active sensors are by physics largely immune to the absence of ambient light. This asymmetry is the core technical argument for multi-sensor stacks in night-dominant use cases.

Section 3 — Tesla’s Camera-Only Approach at Night

Tesla removed radar from new vehicles in 2021 and has never used LiDAR, citing cost and the argument that human-level driving is achievable with human-equivalent sensors (eyes). FSD relies entirely on eight cameras per vehicle — forward narrow and wide, rear, and pillar cameras — plus neural network inference that runs on Tesla’s custom FSD chip.

Dimension	Tesla camera-only at night
Light source dependency	Relies on headlights, streetlights, and other vehicle lights; works in lit suburban and urban environments
Rural and unlit roads	Harder: limited return light; neural network must infer scene geometry from very sparse photons
Oncoming headlight glare	Significant challenge: bright point sources in a dark field can temporarily saturate camera sensors
Wet road reflection	Night rain creates reflections of streetlights and headlights on road surfaces — can confuse lane detection
Neural network compensation	Tesla’s v12 and v13 end-to-end models are trained on billions of night driving clips from the global fleet
Phantom braking at night	Early FSD versions had elevated phantom braking at night (shadows, light patches on road); significantly improved in v12 and v13 (est.)
Tesla’s core argument	Human drivers also use only visible light with no LiDAR; a sufficiently trained vision model can match human night performance

The neural network compensation argument is not without merit. The global Tesla fleet generates an enormous volume of night driving data across diverse lighting conditions — urban arterials, highway onramps, parking structures, rural two-lane roads. End-to-end models trained on this data have, by reported accounts, materially improved phantom braking rates and lane-keeping in variable lighting. The legitimate question is whether learned pattern matching on recorded human night driving reaches the safety floor required for fully driverless commercial operation, or whether it asymptotically approaches but does not surpass the human performance it was trained on.

Section 4 — Waymo’s Multi-Sensor Approach at Night

Waymo uses LiDAR as its primary sensor, supplemented by cameras and radar. The sensor composition at night creates a structural advantage that no amount of neural network training on camera data alone can fully replicate, because LiDAR’s night performance is not a learned capability — it is a physical property of active laser ranging.

Dimension	Waymo multi-sensor at night
LiDAR at night	Operates identically to daytime; full 360-degree point cloud independent of ambient light levels
Active illumination range	LiDAR laser pulses detect pedestrians in near-total darkness at 200-plus meter range (est.)
Camera role at night	Secondary: provides color and texture confirmation of LiDAR-detected objects; performance degrades but LiDAR remains primary
Radar role at night	Velocity measurement of other vehicles; penetrates light rain and fog where cameras degrade
Sensor redundancy	If one sensor degrades (heavy rain on camera lens), other sensors maintain scene representation
Night operational record	Waymo has conducted extensive nighttime driverless operations in Phoenix, including dark desert road segments outside lit urban corridors
Current limitation	Heavy rain and snow still degrade LiDAR range significantly; Waymo has conservatively avoided operating in severe-weather geographies

The sensor redundancy point is architecturally critical. A camera-only system has a single failure mode in low light: the camera sees less. A multi-sensor system has fault tolerance: if the camera lens is rain-soaked, LiDAR and radar continue providing scene representation. If LiDAR return density drops in heavy precipitation, cameras and radar continue providing object detection. The degradation curve for a multi-sensor stack is shallower and more predictable than for a single-modality system, even when individual sensors degrade faster than a camera alone would in light conditions.

Section 5 — Adverse Weather: Rain, Fog, and Snow

Precipitation introduces failure modes that are partially distinct from pure darkness. Each weather type affects the sensor stack differently.

Condition	Tesla camera-only	Waymo LiDAR plus camera plus radar
Light rain	Manageable: cameras still function; wiper-cleared lens maintains most capability	LiDAR mildly attenuated; radar unaffected; overall performance good
Heavy rain	Lens soaking reduces contrast; road reflections increase significantly; capability degraded	LiDAR range reduced; camera degraded; radar compensates for velocity sensing; functional but limited
Dense fog	Camera contrast very low; effective sensing range drops sharply	LiDAR degraded significantly (laser scatter in water droplets); camera degraded; radar helps with moving objects; weather-limited overall
Light snow	Camera workable; white-on-white conditions make lane detection harder	LiDAR can accumulate snow on sensor head (hardware failure mode on spinning units); camera also harder; radar performs best
Heavy snow or accumulation	Lane markings buried; camera largely blind; not operationally safe	All sensors degrade; LiDAR accumulation is a known hardware problem; Waymo avoids snow-belt cities for this reason
Geographic implication	Neither stack is cleared for snow-belt mass deployment as of mid-2026

The fog scenario is particularly instructive. Fog scatters both visible light (degrading cameras) and laser pulses (degrading LiDAR). In dense fog, the multi-sensor advantage narrows substantially — radar is the only sensor that performs reliably, and millimeter-wave radar’s low spatial resolution limits its ability to support autonomous navigation alone. This is the condition in which both architectures converge toward the same constraint: insufficient sensor return for safe high-speed navigation.

Section 6 — What to Watch: Night and Weather as Commercial Ramp Gates

Both Tesla and Waymo operate within conservative operational design domains (ODDs) that exclude their hardest conditions. Understanding what must change to expand those ODDs reveals the near-term investment and technology signals to watch.

Waymo’s expansion path:

Solid-state LiDAR improvements reduce the moving-parts failure mode that makes spinning LiDAR heads vulnerable to snow accumulation. Waymo’s Jaguar I-PACE fleet (Gen 5) and the purpose-built Zeekr vehicle (Gen 6) incorporate sensor packaging improvements that address some thermal and precipitation exposure. Sensor head heating systems — already used in some Waymo vehicles — reduce but do not eliminate snow accumulation risk. Meaningful snow-belt deployment would require a step change in solid-state LiDAR cost and precipitation tolerance.

Tesla’s expansion path:

The v12 and v13 end-to-end model generation represent a genuine architectural shift away from object detection rules toward imitation learning over the full driving distribution. This approach benefits more directly from fleet night data scale than rule-based predecessors did. The remaining open question is whether camera-only training can eventually generalize to lighting conditions and road environments that are underrepresented in the training set, or whether low-frequency edge cases (zero-ambient-light rural roads, black ice under sodium-vapor streetlights, simultaneous glare and wet road reflection) require sensor modalities that cameras cannot substitute. Tesla has reconsidered radar in some internal research contexts (est.); whether this translates to a production hardware change remains unconfirmed.

Shared regulatory frontier:

Regulatory frameworks currently distinguish between geofenced good-weather deployment and general AV operations implicitly rather than explicitly. As both Waymo and Tesla push toward wider commercial deployment, regulators will be forced to specify performance floors for adverse weather and night conditions as a precondition for non-geofenced certification. Companies with demonstrated night and rain operational data will have a structural advantage in those regulatory conversations — Waymo’s Phoenix nighttime operational record is directly relevant evidence; Tesla’s disengagement data in adverse conditions, not publicly disclosed in comparable granularity, is the gap.

Sources: NHTSA Traffic Safety Facts — Fatality Analysis Reporting System (nhtsa.gov); FHWA — Road Weather Management, weather and road safety statistics (ops.fhwa.dot.gov); Tesla Vision camera-only AV stack overview (tesla.com/autopilot); Waymo sensor stack and operational updates (waymo.com/blog/). All figures marked (est.) are estimates based on published research, public operational disclosures, and industry reporting; they have not been independently verified under controlled test conditions and should be treated as directional rather than precise.